Data Generative And Adversarial Research Based On Deep Network

There is no doubt that data are the most important thing in today's information society.With the rapid development of machine learning and artificial intelligence,many research fields have achieved good research results with deep neural networks.Data generative and adversarial research have also made great strides by relying on deep convolutional neural networks.In data generative research,compressed sensing combined with convolutional neural networks is the focus of many researchers.Compressed sensing is a hot research direction.In recent years,many researchers have combined neural networks with compressed sensing.In data adversarial research,many researchers use adversarial learning combined with deep convolutional neural networks to study the privacy protection of data.Based on the deep convolutional neural network,this paper does two things in data generative and adversarial research:First,for data generative research,we combine compressed sensing with deep convolutional neural networks based on image data.We designed a deep neural network model LDIT_CSnet for adaptive image compression reconstruction under single sampling rate,and added the idea of residual learning denoising in deep neural networks.In our network model,the compressed signal value is used not only to obtain the initial reconstruction,but also to improve the denoising accuracy in the depth network iteration process.In existing research,many researchers have limited the diversity of model sampling rates in order to obtain the adaptiveness of the sampling layer.Therefore,based on LDT_CSnet,we propose a multi-sampling rate model Multi LDIT_CSnet for adaptive image compression reconstruction under multi-sampling rate.Multi LDIT_CSnet uses a common depth iteration to fuse multiple sampling rate channels in the network model.Finally,we designed a large number of experiments and existing research to compare the reconstruction effects.The proposed model,whether it is a single sampling rate model or a multi-sampling rate model,has further improved the accuracy of data reconstruction.Secondly,for data adversarial research,we combine the adversarial learning and deep convolutional neural networks based on WIFI data.Data adversarial research solves the privacy protection work of WIFI data in the case of multi-task classification.We first define the multi-task classification scenario of WIFI data,and then explore the importance of WIFI data privacy protection in the actual scenario.Based on the deep convolutional neural network,we designed a WIFI-ATN adversarial model to achieve the protection against sample protection of WIFI data,and set two types of unbiased target and biased target for the protection type.Unbiased target protection means that WIFI data will hide the correct classification result after WIFI-ATN protection.Biased target protection means that WIFI data will not only hide the correct classification result but also predict a certain misclassification result after WIFI-ATN protection.Then,according to these two types of protection,we designed the training scheme,loss function and data processing of the WIFI-ATN model in detail.Finally,we designed a large number of experiments to verify that WIFI-ATN has a good privacy protection effect in the WIFI data adversarial learning.